@xiaogaifun: It seems many people haven't noticed this paper from Anthropic. Anthropic, as a company, does many things with meticulous attention to detail and rigor. Today is Friday, and with nothing much to do in the afternoon, I read through this paper they just published in its entirety. The entire 57 pages systematically explore what would happen if AI capabilities continue to improve rapidly over the next few years and really start...

X AI KOLs Timeline Papers
ai-impact economics job-market automation income-distribution gdp-growth anthropic-research

Summary

This article discusses a paper published by Anthropic, which systematically examines the impact of AI capabilities reaching 2030 on GDP, wages, employment, and income distribution, and analyzes changes under different scenarios.

It seems many people haven't noticed this paper from Anthropic. Anthropic, as a company, does many things with meticulous attention to detail and rigor. Today is Friday, and with nothing much to do in the afternoon, I read through this paper they just published in its entirety. The entire 57 pages systematically explore what would happen if AI capabilities continue to improve rapidly over the next few years and really start to enter the workforce at scale, and by 2030, how GDP, wages, employment, and the entire society's income distribution would change. This analytical framework is very well done, and I gained a lot from reading it. It designs several different AI development scenarios and then looks at the economic and employment impacts in each case. Below are my notes. If you have time, I strongly recommend taking a look, mainly to understand their analytical logic. At the same time, while reading this, you can also intuitively feel the potential impact of AI. 1. The paper roughly divides work into two categories. One is cognitive work, including management, professionals, sales, office work, and so on. Programmers, lawyers, and researchers are largely in this category. The other is other types of work, such as construction, electricians, transportation, and on-site services. They first assume that before 2030, AI will primarily directly affect the first category of work, without considering the rapid entry of robots into manual labor for now. Therefore, the later discussion of AI's impact on work in the paper mainly focuses on today's knowledge work and office work. 2. The paper designs three scenarios: mild, significant, and extreme. In the mild scenario, AI development and adoption are relatively slow. By 2030, GDP is only 1.6% higher than in a no-AI case. In the significant scenario, AI has become a new technology with greater impact than the internet. By 2030, GDP is 8.3% higher than in a no-AI case. In the extreme scenario, AI begins to massively complete cognitive tasks that were previously done by humans. By 2030, GDP is 32.4% higher than in a no-AI case. Note that these three numbers are not the authors' predictions for 2030. They also don't say which scenario is most likely. These are all assumptions. 3. AI's impact on work mainly falls into two situations. One is that AI helps people improve efficiency. For example, programmers, lawyers, and salespeople using AI can complete more work in the same time, and people remain in their original positions. The other is when AI can independently complete tasks. In this case, people start to exit, and the wages that were previously paid to humans will increasingly shift to capital such as models and computing power. 4. The paper argues that the biggest variable affecting AI's impact on employment and income distribution is the automation ratio. The more tasks AI participates in, and the higher the proportion of tasks completed independently by AI, the less work humans need to do, and more income flows to capital. Take the significant scenario mentioned earlier. By 2030, AI can affect about 30% of tasks in the entire economy, but only 40% of those are actually deployed in work. Calculating, about 12% of economic tasks are truly using AI, equivalent to about one-fifth of today's cognitive work tasks. More critically, among these tasks using AI, three-quarters can be completed independently by AI, with only one-quarter still being AI-assisted by humans. In this case, AI will increase the productivity of these tasks by about 57%. The entire economy also accelerates significantly, with GDP 8.3% higher than in a no-AI case by 2030, and annual economic growth reaching 5.4%. This growth rate is already very fast. During the peak of the U.S. internet boom in 1999, real GDP growth was only 4.7%. 5. But after economic growth, these new incomes do not distribute evenly between labor and capital. In the significant scenario, by 2030, GDP is 8.3% higher than in a no-AI case, but average wages are only 2.1% higher. Breaking it down further, the differences are more pronounced. Wages for cognitive work are actually 0.3% lower, while wages for other professions are 5.9% higher. The total income of all laborers combined is only 1.4% higher than in a no-AI case, but capital income is 18.9% higher. The reason lies in the automation discussed earlier. In this scenario, among tasks using AI, three-quarters can be completed independently by AI. Part of the work that previously required paying wages to humans is now handed over to models, computing power, and other capital, with corresponding income flowing more to capital. So although the entire economy grows by 8.3%, the share of labor income in GDP drops from 60% to 56.1%, and the share of capital income rises from 40% to 43.9%. This is also a very important conclusion of the paper. AI can make the entire economy grow very fast, but how much GDP grows and how much ordinary workers' incomes grow are two different things. 6. To be more specific, assume the original GDP is 100, and laborers take 60 of it. Now GDP grows to 132, but laborers can only take 45% of it. 132 times 45% is still about 60. In other words, the entire economy has grown by one-third, but the total income of all laborers combined has almost not increased. The newly generated income has mostly gone to the capital side (AI companies). 7. AI will also revalue work that temporarily cannot be replaced. In the extreme scenario, wages for cognitive work are 11.5% lower than in a no-AI case, but wages for other professions are 33.6% higher. The logic behind this is that AI reduces the cost of many cognitive tasks, and the output of the entire economy increases rapidly as a result. But for the economy to produce so much more, it still requires work like construction, electricians, transportation, and on-site services that AI temporarily cannot handle. So an interesting change occurs. On the cognitive work side, AI can directly supplement supply, reducing demand for humans and suppressing wages. On the other work side, AI temporarily cannot supplement, but after the entire economy expands, demand for these jobs increases instead. Even if some cognitive workers switch over, they may not fully meet the new demand, so wages rise. This is also why as AI becomes stronger, wages for different professions do not necessarily rise or fall together. The easier work is for AI to complete, the more human value is suppressed; the harder work is for AI to take over, the more it might become more valuable due to overall economic growth. 8. As mentioned earlier, AI will reduce demand for some cognitive work, while overall economic growth will increase demand for work like construction, electricians, transportation, and on-site services. The problem is that people cannot immediately switch careers following job changes. The paper gives a vivid example. A software engineer who loses their job may find it hard to immediately become an electrician. Switching industries requires relearning, job searching, and past experience may not be directly applicable. Of course, it doesn't have to be an electrician; it could be an insurance manager or other service industry jobs. So even if there are new job opportunities in the overall economy, unemployment can still occur in between. The faster AI takes over old jobs, and the slower laborers switch professions, the more people will temporarily be unable to find new work. 9. The same AI shock can result in two outcomes: wage reduction or unemployment. The paper makes a specific comparison. With AI capabilities and automation levels held constant, it only looks at whether wages can decrease when companies need fewer workers. If wages can decrease quickly, companies will continue to retain more people. The result is that cognitive workers' wages are 42.2% lower than in a no-AI case, but the unemployment rate is only 2.6%. If wages are hard to decrease, companies need fewer workers, but wages don't fall, so they can only hire fewer and lay off more. In this case, cognitive workers' wages might even be 2.8% higher than in a no-AI case, but the unemployment rate rises to 24%. So after AI reduces labor demand, the outcome does not necessarily manifest only as unemployment. The easier wages are to decrease, the more impact falls on income. The harder wages are to decrease, the more impact falls on employment. 10. There's another concept I find very insightful, called so-so automation, which can be understood as automation with low gains. For example, if a person used to spend 100 dollars to complete a task, now AI can take over the task, reducing the cost to 95. For the company, humans can indeed be eliminated. But from society's perspective, this technology only creates a 5-dollar efficiency improvement. The large chunk of income previously paid to workers has shifted to the capital side along with the task. This kind of AI will replace many jobs, but society doesn't create much additional wealth as a result. 11. The more capable AI is of independently completing tasks, the more companies need additional computing power to handle these tasks. If computing power supply can quickly keep up, companies need more computing power, the market can increase it in time, and prices won't be significantly pushed up. In this way, capital gains from scarcity won't be too large, and laborers can share more through wages. If computing power supply cannot keep up, the situation reverses. All companies are vying for limited computing power and equipment, and the higher the demand, the easier it is for prices to rise, with more gains flowing to capital owners, naturally reducing the portion left for wages. Of course, this model also makes many simplifications. It puts computing power into a unified capital category without separately accounting for changes like falling model prices and inference costs. In reality, these factors might alleviate some capital scarcity. But if in the future, models have very strong long-horizon task capabilities, and AI truly starts to take over work at scale, then a very practical question is whether there is enough computing power, data centers, and electricity to actually run these capabilities. At that stage, computing power supply is likely to become a major bottleneck. 12. However, even in the most extreme scenario, the new gains created by AI are still far higher than the income lost by cognitive workers. The paper calculates that the gains for the entire economy are about three times the income loss of cognitive workers. Theoretically, by allocating income equivalent to 9% of GDP, affected cognitive workers could maintain their original income levels. For the rest, overall income could still be more than 20% higher than in a no-AI case. So the real difficulty may not be that AI doesn't create enough wealth, but how this wealth is ultimately distributed. The paper also specifically mentions that in the past, whether it was technological change or trade shocks, beneficiaries rarely took the initiative to compensate those affected. If AI truly reaches the extreme scenario, the issue to solve may be this: how to ensure those who are replaced by AI or have their wages reduced also share in the growth created by AI. 13. The path of AI accelerating scientific research, surprisingly, does not contribute much to economic growth before 2030. Even in the extreme scenario, AI can indeed help scientific research, but much research is ultimately limited by experiments and physical world factors. So they calculate that the additional labor productivity gains from accelerating research are less than 1% in all three scenarios. Of course, the authors themselves admit that this part of the model does not fully account for the recursive feedback of AI helping AI research, so this result is likely conservative. It's quite interesting, and I recommend everyone to take a look. Anthropic has also made a beautiful analysis webpage, and I've put the paper link in the comments.
Original Article
View Cached Full Text

Cached at: 09/18/26, 10:48 PM

It seems many people missed this paper from Anthropic.

Anthropic is indeed a company that does many things meticulously and rigorously. Today is Friday afternoon, and since there’s not much else to do, I read through the entire paper they just published.

The full 57-page document systematically examines what might happen to GDP, wages, employment, and overall income distribution by 2030 if AI capabilities continue to advance rapidly in the coming years and truly begin to enter the workforce at scale.

The analytical framework presented here is exceptionally well-designed—personally, I found it quite insightful after going through it. The paper outlines several different scenarios for AI development and then assesses the economic and employment impacts under each. Below are my notes.

If you have time, I strongly recommend giving it a read—primarily for their analytical approach. It also offers a fairly intuitive sense of the potential impacts AI might bring.

  1. Work Categories
    The paper roughly divides work into two types:
  • Cognitive work, including management, professional roles, sales, office jobs, etc. Programmers, lawyers, and researchers largely fall into this category.
  • Other work, such as construction, electrical work, transportation, on-site services, and similar roles.

The study assumes that before 2030, AI will primarily impact cognitive jobs directly, without yet considering the rapid integration of robots into physical labor.

Therefore, the later discussion on AI’s disruption of work mainly centers on today’s knowledge-based and office occupations.

  1. Three Scenarios
    The paper outlines three scenarios: Mild, Significant, and Extreme.
  • In the Mild scenario, AI develops and spreads relatively slowly. By 2030, GDP would be only 1.6% higher than in a world without AI.
  • In the Significant scenario, AI becomes a transformative technology with even broader impact than the internet. By 2030, GDP would be 8.3% higher than without AI.
  • In the Extreme scenario, AI begins to autonomously handle large portions of cognitive work. By 2030, GDP would be 32.4% higher than without AI.

Note that these three figures are not the authors’ predictions for 2030, nor do they assign probabilities to any scenario. They are hypothetical models.

  1. Two Modes of AI Impact
    AI affects work in two main ways:
  • AI augmenting humans: For example, programmers, lawyers, or salespeople use AI to accomplish more work in the same time, while remaining in their original roles.
  • AI performing tasks independently: In this case, humans start to exit, and the wages previously paid to them increasingly shift toward capital—models, computing power, etc.
  1. Key Variable: Automation Ratio
    The paper argues that the most critical variable affecting employment and income distribution is the proportion of tasks automated. The more tasks AI participates in—and especially the higher the share completed autonomously—the less human labor is required, and the more income flows to capital.

Take the Significant scenario again:
By 2030, AI could affect about 30% of tasks across the economy, but only 40% of that would be deployed in actual jobs.
That means roughly 12% of economic tasks truly utilize AI—about one-fifth of today’s cognitive work.
More crucially, of those tasks using AI, three-quarters could be completed autonomously by AI, with only one-quarter still involving AI assisting humans.

In this scenario, AI raises productivity for these tasks by around 57%. The overall economy also accelerates noticeably, with GDP in 2030 being 8.3% higher than without AI, and that year’s economic growth reaching 5.4%.

That growth rate is already quite fast. During the peak of the U.S. internet boom in 1999, real GDP growth was only 4.7%.

  1. Income Distribution Disparity
    However, after economic growth, the new income doesn’t necessarily reach workers and capital equally.

In the Significant scenario, while GDP is 8.3% higher than without AI, average wages are only 2.1% higher.
Breaking it down further reveals a starker contrast:

  • Wages for cognitive work actually decrease by 0.3%, while wages in other occupations rise by 5.9%.
  • Total labor income across all workers increases by only 1.4% compared to the no-AI case, but capital income surges by 18.9%.

The reason lies in the automation mentioned earlier.
In this scenario, three-quarters of AI-involved tasks can be completed independently by AI.
Work that previously required human labor—and thus human wages—is increasingly handled by models, computing power, and other capital. The corresponding income flows more toward capital.

Therefore, even though overall GDP grows by 8.3%, labor’s share of GDP declines from 60% to 56.1%, while capital’s share rises from 40% to 43.9%.

This is a crucial conclusion of the paper: AI can drive rapid overall economic growth, but GDP growth and wage growth for ordinary workers are two different things.

  1. A Concrete Example
    Suppose original GDP was 100, with labor receiving 60.
    Now GDP increases to 132, but labor’s share drops to 45%.
    132 × 45% = approximately 59.4—still roughly 60.

So even as the economy expands by a third, total income for all workers barely increases. The majority of new gains go to capital (i.e., AI companies).

  1. Revaluation of Non-Automatable Jobs
    AI will also raise the value of jobs that are temporarily hard to automate.
    In the Extreme scenario, wages for cognitive work drop by 11.5% compared to the no-AI case, but wages in other occupations rise by 33.6%.

The logic behind this is that AI drastically reduces the cost of many cognitive tasks, boosting overall economic output.
However, to produce all this additional output, the economy still needs jobs in construction, electrical work, transportation, and on-site services—roles AI cannot yet handle.

This creates an interesting dynamic:
On the cognitive side, AI directly increases supply, reducing human demand and depressing wages.
On the other side, AI can’t yet fill the gap, but economic expansion increases demand for these roles. Even if some cognitive workers switch fields, they may not fully meet the new demand, pushing wages up.

This also explains why AI advancement doesn’t necessarily lift all occupational wages simultaneously.
The easier a job is for AI to perform, the more human value in that role gets suppressed. The harder it is for AI to take over, the more valuable those roles may become due to broader economic growth.

  1. Transition Challenges
    As mentioned, AI reduces demand for some cognitive jobs, while overall economic growth increases demand for roles like construction, transportation, and on-site services.
    But people can’t instantly switch careers when jobs change.

The paper gives a vivid example: a software engineer who loses their job can’t immediately become an electrician. Changing industries requires re-learning, job searching, and much of their past experience may not directly transfer.

It could be an electrician or a service industry role like an insurance manager.

So even if the economy creates new job opportunities, there can still be unemployment in the transition. The faster AI takes over old jobs and the slower workers switch, the larger this temporary group of unemployed individuals will become.

  1. Wage Flexibility vs. Employment Impact
    The same AI shock can lead to two outcomes: either reduced wages or unemployment.
    The paper compares two sub-scenarios:
  • If wages can adjust downward quickly, companies retain more workers. Cognitive wages fall by 42.2% compared to the no-AI case, but unemployment is only 2.6%.
  • If wages are sticky and don’t fall, companies needing fewer workers will simply hire less and lay off more. In this case, cognitive wages might even rise by 2.8%, but unemployment soars to 24%.

So after AI reduces labor demand, the impact isn’t necessarily all unemployment. If wages can fall easily, the impact leans more toward income reduction. If wages are hard to cut, the impact shifts more toward job losses.

  1. “So-So Automation”
    Another insightful concept in the paper is “so-so automation.”
    For example, if a task used to cost 100 in labor, and AI now handles it for 95, the company can replace the worker. But from society’s perspective, efficiency only improves by 5.

The large portion of income previously paid to labor has already moved to capital with the task. This type of AI might replace many jobs without creating proportionally more wealth for society.

  1. Compute as a Potential Bottleneck
    The more independently AI can perform tasks, the more computing power companies need to handle them.
    If compute supply grows quickly enough to meet demand, prices won’t spike significantly. Capital won’t capture excessive rents due to scarcity, and workers could still gain more through wages.
    If compute supply lags, the reverse happens: companies scramble for limited compute and infrastructure, driving up prices and shifting even more gains to capital owners, leaving less for wages.

The paper simplifies by grouping compute into a single capital category and doesn’t model continuous declines in model pricing and inference costs. In reality, these factors might mitigate some capital scarcity.

But if AI’s long-horizon capabilities become truly strong and large-scale job integration begins, a pressing practical question emerges: Is there enough compute, data centers, and electricity to actually run these systems?

At that stage, compute supply could become a critical bottleneck.

  1. The Distribution Challenge
    Even in the most extreme scenario, the gains generated by AI far exceed the income losses for cognitive workers.
    The paper calculates that total economic gains are roughly three times the income loss for cognitive workers.
    In theory, dedicating about 9% of GDP in revenue could restore affected cognitive workers’ original income levels. Everyone else would still see overall income rise by over 20% compared to the no-AI case.

So the real issue may not be insufficient wealth creation by AI, but rather how that wealth is distributed.
The paper notes that historically, whether with technological change or trade shocks, beneficiaries rarely voluntarily compensate those negatively affected.
If AI truly reaches an extreme scenario, the core challenge may be how to ensure those displaced or whose wages are reduced also share in AI-driven growth.

  1. Limited Impact from Accelerated Scientific Research
    The pathway of AI accelerating scientific research actually doesn’t contribute significantly to economic growth before 2030.
    Even in the extreme scenario, while AI can assist research, many studies still face constraints from experiments and the physical world.
    The paper estimates that additional labor productivity gains from accelerated research remain below 1% across all three scenarios.

The authors themselves acknowledge the model doesn’t fully capture the recursive feedback of AI helping AI research, so this result may be conservative.

It’s a fascinating read—I recommend checking it out. Anthropic also created a clean explanatory webpage for the paper; I’ll leave the link in the comments.

Similar Articles

@ba_niu80557: https://x.com/ba_niu80557/status/2071277244287426980

X AI KOLs Timeline

The article deeply analyzes the internal changes Anthropic faces as AI-generated code becomes extremely efficient: the bottleneck shifts from 'writing' to 'verification', traditional management, long-term planning, and effort measurement become ineffective, attention becomes the new scarce resource, and engineers even feel lonely. These phenomena foreshadow the challenges other companies may face in the future.